Carry 3D Gaussian splats, made from point clouds, in ordinary z/x/y map tiles — and draw them in MapLibre GL JS.
An experiment, not a standard. 日本語の要約は下
Try it in your browser → (needs WebGL2) · a real point cloud at Numazu, Japan · a bridge with its underside · an overhang
Drag to pan, right-drag or Ctrl-drag to rotate and tilt. The camera is in the URL, so you can share a view.
Horizontal addressing uses tiles. Tile contents use surfels.
Map tiles address the world in two dimensions. Terrain in a tile is usually
a height field: one height per horizontal position. That cannot hold a
bridge and the road under it, a vertical wall, an overhang, or a tree
crown. Dorayaki keeps the familiar z/x/y addressing and puts a
three-dimensional distribution of splats inside each tile. Several splats
may share the same horizontal position at different heights and
orientations.
A splat is the primitive of 3D Gaussian Splatting: a position, three scales, a rotation and a colour. Here it is quantized into 16 bytes.
- Round trip. A Surfel Block (the tile's contents) encodes and decodes
within stated quantization tolerances (
npm test). - Not a height field. The
bridgeandoverhangfixtures hold multiple heights at the same x/y and render correctly from above and below. - Real data, small. One mobile-mapping point cloud mesh
(976,195 points, VIRTUAL SHIZUOKA, Numazu) becomes a 603 KB PMTiles
archive in its default
coarseform: 0.12× the size of the same points as COPC. That is a coarse picture (one splat per 12 cm cell), not a lossless copy; finer profiles are 0.43× and 0.50× COPC. See documents/numazu-demo.md and EXPERIMENTS.md. - Real data, large. The Kawazu loop bridge and the valley below it: 476 million points (mobile mapping and airborne laser, 1.08 km2, nine survey meshes) become an 83 MB PMTiles archive of 6.94 million splats (about 77 bytes per square metre), served from stars. The demo loads only the tiles in view. Vegetation is kept as coarse blobs; flat, evenly coloured ground as large splats. See documents/kawazu-demo.md, documents/large-point-clouds.md and documents/size-and-scale.md.
- Ordinary delivery. The Numazu tiles are served from a normal HTTP tile endpoint (TileJSON, backed by PMTiles) and drawn by a MapLibre GL JS custom layer, with pan, zoom, bearing and pitch.
- A point cloud is fitted with one Gaussian per voxel (
src/pointcloud/). - Gaussians are assigned to the one tile containing their centre. Zoom levels below the maximum hold coarser aggregates; the level of detail is the ordinary tile zoom hierarchy.
- Each tile's splats are packed into a Surfel Block (
DSRF, version 5 or 6; the hosted Kawazu archive is version 6, 12 bytes per splat) and wrapped in an experimental MapLibre Tile frame. - The browser fetches
z/x/y, unwraps the frame, decodes the block, and draws instanced ellipses (round splats face the camera).
The three concerns are separable: the Surfel Block codec runs and is tested without MapLibre; the frame wrapper skips unknown frames and rejects malformed lengths; the renderer only sees decoded splats.
Please read this before depending on anything here.
- Not a specification. The byte layout changes without notice and version numbers rise freely (it is at version 6, opt-in, and 5; breaking changes are stated in the format document).
- Not an MLT assignment. Frame tag
0xD0is a local, unassigned value chosen for the experiment. The upstream MapLibre Tile implementation is unmodified. - Two real datasets, one valley. Numazu (one mesh) and Kawazu (nine meshes, 1.08 km2). Size figures for a whole prefecture are extrapolations from Kawazu alone.
- Tile selection follows map zoom, not the camera's distance to the data, so tilted views at close range can pick the wrong detail.
- Frame time was 16 to 17 ms (the display rate) at the Kawazu views tried on one Apple M1, which says little about weaker GPUs. Browser decoding measured about 0.23 µs per splat on a warm 389,000-splat tile (E011, size-and-scale.md); a first, cold call on a small tile was about 3 µs (NOTES.md).
- Mercator only: no globe. Terrain is needed for data far above sea
level (the camera is otherwise relative to 0 m): the Kawazu demo turns it
on by default (lowered by 2 m); elsewhere it is the
?terrain=1flag. Heights are orthometric (標高); the geoid model is not recorded in the block. - No compression beyond quantization, no spherical harmonics, no source-point storage.
Observed findings are logged in NOTES.md; measured experiments in EXPERIMENTS.md.
| document | contents |
|---|---|
| documents/surfel-block-format.md | byte layout, quantization, coordinate frame |
| documents/mlt-frame.md | the experimental MLT frame envelope |
| documents/progressive-maxzoom.md | maxzoom and nested refinement order |
| documents/numazu-demo.md | rebuilding and hosting the Numazu tiles; profiles |
| documents/demo-and-site.md | running the demo locally, URL flags, demo modules, the Pages site |
| documents/large-point-clouds.md | streaming build of big LAS files; Virtual Shizuoka data and Kawazu |
| documents/adaptive-merging.md | merging coplanar splats; covering gaps (?extent=) |
| documents/size-and-scale.md | what a tile and a prefecture cost; browser memory |
| documents/kawazu-demo.md | the Kawazu tileset: recipe, size, hosting |
| AGENTS.md | project goals and working method |
Node.js 22 or later.
npm install
npm run typecheck # type-check src/, test/, scripts/ and demo/src/
npm test # compile and run node --test
npm run build # compile src/ into dist/
npm run demo # generate tiles, bundle, serve the demo locallynpm run demo serves on 127.0.0.1:8000; details and the other scripts
(demo:numazu, site, measure:*) are in
documents/demo-and-site.md and AGENTS.md.
Source layout: src/surfel-block.ts (codec), src/mlt-frame.ts (frame),
src/pointcloud/ (point cloud → Gaussians → tile pyramid),
src/fixtures/ (deterministic synthetic data), demo/ (browser demo),
test/.
点群から作った 3D Gaussian Splatting のスプラットを、通常の z/x/y 地図タイルに
載せて MapLibre GL JS で描画する実験です。標準仕様ではありません。タイルの
住所は従来どおり2次元、中身は3次元のスプラットなので、橋とその下の道路、垂直面、
オーバーハング、樹冠のように、同じ水平位置に複数の高さがあるものも保持できます。
沼津の MMS 点群(976,195点)は、既定の coarse で 603 KB の PMTiles
(同じ点の COPC の 0.12 倍。粗い表現であり可逆ではありません)になります。
河津ループ橋と谷(1.08 km²、4.76 億点、航空レーザーと MMS)は、83 MB の PMTiles
(約 77 B/m²)で、stars から配信しています。表示範囲のタイルだけを読み込みます。
サイズの県全体への換算は、この谷 1 か所からの外挿です。
ブラウザで試す。
CC0 1.0 Universal. See LICENSE.